Energy characterization and optimization of cloud datacenters

Cloud datacenter energy consumption has grown in recent years. In particular, the CPU is the most power hungry components in the datacenter because they are not energy proportional to their utilization. A cloud server's energy efficiency is much lower as the utilization of the CPU gets closer to idle. At the same time, current cloud computing applications usually have significant CPU idle times composted of short and variable idle intervals. The power consumption in these idle intervals is significant due to leakage power prominent in recent semiconductor technologies. Therefore, in this paper we present the characterization of a potential workload considering several factors such as quality of service, service level agreements, host utilization policies and virtual machine migration policies. We also present a study of a number of schemes that transition the CPU into various low power states and sleep states to reduce the CPU idle power. The disadvantage of utilizing a sleep state is the potential negative power savings if its wakeup latency is longer than the current idle interval. This makes finding an intelligent sleep state entry the key to improving datacenter CPU energy usage. In this paper, we propose a dynamic idle interval prediction scheme that can forecast the current CPU idle interval length and thereby choose the most cost-effective sleep state for achieving the minimized power consumption. The proposed approach largely outperforms other schemes examined, achieving better power savings compared to DVFS when using various CPU idle patterns.